Runway says its AI-led software workflow cut approval waits to five minutes
The company’s internal deployment extends beyond writing code to testing, fixing failures and assigning work. High-risk changes still require human approval.
Loading page…
The company’s internal deployment extends beyond writing code to testing, fixing failures and assigning work. High-risk changes still require human approval.
Listen to this story
Runway says its internal delivery system has grown from a coding assistant into a coordinated workflow that handles pull requests, browser testing, repairs and failed checks, while people set plans and retain control over sensitive approvals. The system dispatched more than 14,000 tasks in its first two live weeks, producing about 1,900 pull requests; median time to first approval fell from more than two hours in September to about five minutes by mid-November. These are company-reported results, and Runway says the bottleneck is shifting upstream to deciding what to build and answering agents’ questions.
Cloud agents drove 38% of merged pull requests in October, up from 0.1% in June.
Of 100 agent-driven requests merged in early November, 78 had no human code commits; that does not mean they lacked human direction or review.
AI-only approvals are allowed for low-risk changes and some medium-risk changes; high-risk work and designated critical areas still require human sign-off.
At Runway, AI agents can now carry software changes beyond writing code: they review changes, approve some for merging, test the product and respond to failures. In a November 20 engineering disclosure, the company said its median wait for first approval fell from over two hours in September to about five minutes in mid-November.
Runway’s starting problem was a growing review queue. Over the summer, the share of human reviews spent on agent-written code in its core API repository more than tripled, to about one in five. The number of proposed changes awaiting human approval doubled.
Its response was an internal cloud service that stays with each pull request—a proposed change to a codebase—until automated checks pass. A person assigns the task and approves the plan. The agent then writes code, opens the request, fixes failing tests, responds to review comments and updates the change against the latest code.
Cloud agents drove 38% of merged pull requests in October, up from 0.1% in June, Runway said. In a sample of 100 agent-driven requests merged in early November, 78 contained no code commits from a person. That measures hands-on code changes, not the absence of human direction or review.
Next came an AI reviewer supplied with the coding agent’s plan, the proposed edits and surrounding code. The two agents exchange fixes until the reviewer is satisfied. But reviewing a change and authorizing it to merge are separate jobs. Runway began allowing merges on AI approval alone in mid-October, after first running AI approvals alongside people.
A separate agent assigns each request a risk level, choosing the higher category when uncertain. Runway’s risk-based approval rules then determine whose sign-off counts:
Across September and October, people objected to 15 of the AI reviewer’s 977 approvals. Runway also reported fewer reversals after AI-only approvals than human approvals, but noted that AI mostly approved lower-risk work. Its comparison within risk levels found comparable reversal rates. These are Runway’s internal results, not an independent evaluation.
For App Studio, Runway built a testing loop in which agents operate browsers like users. Before filing a bug, an agent checks its own setup and locates the responsible code. Confirmed bugs become repair tasks, followed by another test run. In its first four days, the loop found 86 confirmed bugs; agents fixed 74, with 67 verified by rerunning tests.
When checks fail on the main codebase, agents retry them, diagnose the problem and apply known fixes before calling a person. In its first two weeks, that workflow encountered about 140 alerts. All but one were resolved without calling anyone in, with a median 23 minutes to restore passing checks. Separately, the release system automatically reverses production releases that fail health checks.
A coordinator connects those steps: it reads the issue tracker, assigns agents and orders merges, escalating when policy requires a person or work gets stuck. Runway tested it first without write access, then had people judge proposed actions. Its first two live weeks dispatched over 14,000 agent tasks, leading to about 1,900 pull requests.
With more of delivery automated, Runway says the wait is shifting upstream: deciding what to build and answering agents’ questions.
Loading discussion...
Join the conversation
Explain which safeguards would make that acceptable—or why a person should always review.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.